@@ -861 +861 @@
-        hidden_states = inputs_embeds + position_embeds.to(inputs_embeds.device)
+        hidden_states = inputs_embeds + position_embeds
@@ -867,3 +867 @@
-        causal_mask = self._update_causal_mask(
-            attention_mask, inputs_embeds, cache_position, past_key_values, output_attentions
-        )
+        causal_mask = None
@@ -877 +875 @@
-            if encoder_attention_mask is None:
+            if not _use_sdpa and encoder_attention_mask is None:
@@ -880,3 +878 @@
-                encoder_attention_mask = _prepare_4d_attention_mask_for_sdpa(
-                    mask=encoder_attention_mask, dtype=inputs_embeds.dtype, tgt_len=input_shape[-1]
-                )
+                pass